| Improving urban traffic management and solving traffic congestion effectively has become a key task urgently in big cities,and traffic prediction is regarded as one of the key elements for it.With the development of deep learning theory and neural network model,more and more modeling ideas had inspired many researchers.This paper proposes a short-term prediction model and a long-term prediction model with deep neural network theory of sequential data,which provides fully support for a traffic induction,travel information service and traffic planning and management.Firstly,in view of the traffic prediction problem,this paper has summaried the research background,and has reviewed many related researches,which concludes the idea of discriminative model for short-term prediction and generative model for long-term prediction.And then the research target and technical route is presented,This paper also summaries the basic theories of deep neural network model which is related for this paper.Secondly,according to the quality of traffic data set,this paper explaines the related pre-processing operation.With the statistical characteristics of traffic data,this paper analyzes the time-varying correlation of traffic road network by using correlation indexes,and then constructs quantitative indexes to analyze the spatial heterogeneity of traffic road network.The analysis results provide a theoretical support for traffic prediction modelling.Thirdly,in combination with the short-term traffic prediction and discriminative modelling theory,this paper defines short-term traffic discriminative prediction problem.With the corresponding key points,a multi-step prediction neural network model based on spatiotemporal information aggregation(GCGRUatt)is proposed.As for this model,regarding traffic data of road network as spatio-temporal graph structure,GCGRUatt model merges spectral graph convolution with gated recurrent unit,and use attention mechanism to strengthen the temporal dependence modelling,which has realized the multi-step traffic prediction at the road network level.Analysis results show that the accuracy and stability of multi-step prediction within 60 min are better than other short-term prediction models.Besides,this model has interpretability to some extent.Finally,in combination with the long-term traffic prediction and inferential generative modeling theory,this paper defines the problem of long-term traffic inferential generative prediction.With the corresponding key points,a time series generation neural network model based on state space inference(Sequence VAE)is proposed.This model modifies the structure of multi-layer perceptron to realize the operation of probability distribution estition and sampling with neural network model,and use variational inference to construct the objective function,so that the model can learn the distribution of state space variables,and also the traffic data sequences can be generated by sampling from state space variables.Analysis results show that the 24-hour long-term prediction effects are better than other short-term prediction models,and this model has interpretability to some extent,which further demonstrates the advantages and potential of the inferential generative model in long-term prediction.There are 55 figures,14 tables and 89 references in the body. |
